Unverified50% confidenceFactExact time
千问3.8 Flash Next额外引入了510亿参数的Ngram模块，类似巨大的局部短语记忆表
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified千问3.8 Flash Next是一款采用下一代千问4架构的Flash模型，开源当天放出了完整权重72% similarVerifiedDeepSeek V4 Flash uses a Mixture of Experts (MOE) architecture with a total parameter count exceeding 600 billion (600B+).67% similarUnverifiedFlashAttention由斯坦福大学的Tri Dao等人提出，通过将注意力计算分块在GPU片上SRAM中完成，将内存访问次数降低一个数量级，是数学上精确等价的实现66% similarUnverified盘古2.0 Flash采用MoE（混合专家）架构，参数规模为92B A6B（总参数92B，激活参数6B）65% similarUnverified混元3配备约38亿参数的MTP模块，用于推理加速65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843530API
curl https://kongchang.com/api/v1/knowledge/claims/843530MCP
get_claim(id=843530)